Abstract:
The condenser vacuum is an important variable in steam power plants. Monitoring and controlling this
variable requires predicting its behavior. This paper develops further Autoregressive-Generalized
Autoregressive Conditional Heteroscedasticity (AR-GARCH) models for this purpose, using lagged values
of predictors. The predictors include the inlet temperature of the condenser cooling water and the active
power of the generator. Models can be adequately trained with small-sized data, making them suitable for
use in thermal plants, which are often regularly maintained with operating conditions being reset,
rendering past data obsolete. Training and testing were carried out using operation data from an actual
steam power plant generating unit during a period in which it faced the prospect of an emergency turbine
shutdown. When the models pass all the required statistical tests, they tend to outperform other
techniques, including autoregressive neural networks and support vector regression, in terms of
prediction. This study also discusses an implementation scenario. The choice of training sizes and model
variants can be flexible, enhancing the models' practicality for real operational situations. This study also
provides additional directions for further research.
